tiny-cnn开源库的使用 MNIST

最新推荐文章于 2021-05-05 15:29:23 发布

有点模糊

最新推荐文章于 2021-05-05 15:29:23 发布

阅读量230

点赞数

本文链接：https://blog.csdn.net/hddryjv/article/details/83892128

版权

分享一下我老师大神的人工智能教程！零基础，通俗易懂！http://blog.csdn.net/jiangjunshow

也欢迎大家转载本篇文章。分享知识，造福人民，实现我们中华民族伟大复兴！

tiny-cnn是一个基于CNN的开源库，它的License是BSD 3-Clause。作者也一直在维护更新，对进一步掌握CNN很有帮助，因此下面介绍下tiny-cnn在windows7 64bit vs2013的编译及使用。

1. 从https://github.com/nyanp/tiny-cnn下载源码：

$ git clone https://github.com/nyanp/tiny-cnn.git 版本号为77d80a8，更新日期2016.01.22

2. 源文件中已经包含了vs2013工程，vc/tiny-cnn.sln，默认是win32的，examples/main.cpp需要OpenCV的支持，这里新建一个x64的控制台工程tiny-cnn；

3. 仿照源工程，将相应.h文件加入到新控制台工程中，新加一个test_tiny-cnn.cpp文件；

4. 将examples/mnist中test.cpp和train.cpp文件中的代码复制到test_tiny-cnn.cpp文件中；

#include <iostream>#include <string>#include <vector>#include <algorithm>#include <tiny_cnn/tiny_cnn.h>#include <opencv2/opencv.hpp>using namespace tiny_cnn;using namespace tiny_cnn::activation;// rescale output to 0-100template <typename Activation>double rescale(double x){ Activation a; return 100.0 * (x - a.scale().first) / (a.scale().second - a.scale().first);}void construct_net(network<mse, adagrad>& nn);void train_lenet(std::string data_dir_path);// convert tiny_cnn::image to cv::Mat and resizecv::Mat image2mat(image<>& img);void convert_image(const std::string& imagefilename, double minv, double maxv, int w, int h, vec_t& data);void recognize(const std::string& dictionary, const std::string& filename, int target);int main(){ //train std::string data_path = "D:/Download/MNIST"; train_lenet(data_path); //test std::string model_path = "D:/Download/MNIST/LeNet-weights"; std::string image_path = "D:/Download/MNIST/"; int target[10] = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 }; for (int i = 0; i < 10; i++) {  char ch[15];  sprintf(ch, "%d", i);  std::string str;  str = std::string(ch);  str += ".png";  str = image_path + str;  recognize(model_path, str, target[i]); } std::cout << "ok!" << std::endl; return 0;}void train_lenet(std::string data_dir_path) { // specify loss-function and learning strategy network<mse, adagrad> nn; construct_net(nn); std::cout << "load models..." << std::endl; // load MNIST dataset std::vector<label_t> train_labels, test_labels; std::vector<vec_t> train_images, test_images; parse_mnist_labels(data_dir_path + "/train-labels.idx1-ubyte",  &train_labels); parse_mnist_images(data_dir_path + "/train-images.idx3-ubyte",  &train_images, -1.0, 1.0, 2, 2); parse_mnist_labels(data_dir_path + "/t10k-labels.idx1-ubyte",  &test_labels); parse_mnist_images(data_dir_path + "/t10k-images.idx3-ubyte",  &test_images, -1.0, 1.0, 2, 2); std::cout << "start training" << std::endl; progress_display disp(train_images.size()); timer t; int minibatch_size = 10; int num_epochs = 30; nn.optimizer().alpha *= std::sqrt(minibatch_size); // create callback auto on_enumerate_epoch = [&](){  std::cout << t.elapsed() << "s elapsed." << std::endl;  tiny_cnn::result res = nn.test(test_images, test_labels);  std::cout << res.num_success << "/" << res.num_total << std::endl;  disp.restart(train_images.size());  t.restart(); }; auto on_enumerate_minibatch = [&](){  disp += minibatch_size; }; // training nn.train(train_images, train_labels, minibatch_size, num_epochs,  on_enumerate_minibatch, on_enumerate_epoch); std::cout << "end training." << std::endl; // test and show results nn.test(test_images, test_labels).print_detail(std::cout); // save networks std::ofstream ofs("D:/Download/MNIST/LeNet-weights"); ofs << nn;}void construct_net(network<mse, adagrad>& nn) { // connection table [Y.Lecun, 1998 Table.1]#define O true#define X false static const bool tbl[] = {  O, X, X, X, O, O, O, X, X, O, O, O, O, X, O, O,  O, O, X, X, X, O, O, O, X, X, O, O, O, O, X, O,  O, O, O, X, X, X, O, O, O, X, X, O, X, O, O, O,  X, O, O, O, X, X, O, O, O, O, X, X, O, X, O, O,  X, X, O, O, O, X, X, O, O, O, O, X, O, O, X, O,  X, X, X, O, O, O, X, X, O, O, O, O, X, O, O, O };#undef O#undef X // construct nets nn << convolutional_layer<tan_h>(32, 32, 5, 1, 6)  // C1, 1@32x32-in, 6@28x28-out  << average_pooling_layer<tan_h>(28, 28, 6, 2)   // S2, 6@28x28-in, 6@14x14-out  << convolutional_layer<tan_h>(14, 14, 5, 6, 16,  connection_table(tbl, 6, 16))              // C3, 6@14x14-in, 16@10x10-in  << average_pooling_layer<tan_h>(10, 10, 16, 2)  // S4, 16@10x10-in, 16@5x5-out  << convolutional_layer<tan_h>(5, 5, 5, 16, 120) // C5, 16@5x5-in, 120@1x1-out  << fully_connected_layer<tan_h>(120, 10);       // F6, 120-in, 10-out}void recognize(const std::string& dictionary, const std::string& filename, int target) { network<mse, adagrad> nn; construct_net(nn); // load nets std::ifstream ifs(dictionary.c_str()); ifs >> nn; // convert imagefile to vec_t vec_t data; convert_image(filename, -1.0, 1.0, 32, 32, data); // recognize auto res = nn.predict(data); std::vector<std::pair<double, int> > scores; // sort & print top-3 for (int i = 0; i < 10; i++)  scores.emplace_back(rescale<tan_h>(res[i]), i); std::sort(scores.begin(), scores.end(), std::greater<std::pair<double, int>>()); for (int i = 0; i < 3; i++)  std::cout << scores[i].second << "," << scores[i].first << std::endl; std::cout << "the actual digit is: " << scores[0].second << ", correct digit is: "<<target<<std::endl; // visualize outputs of each layer //for (size_t i = 0; i < nn.depth(); i++) { // auto out_img = nn[i]->output_to_image(); // cv::imshow("layer:" + std::to_string(i), image2mat(out_img)); //}  visualize filter shape of first convolutional layer //auto weight = nn.at<convolutional_layer<tan_h>>(0).weight_to_image(); //cv::imshow("weights:", image2mat(weight)); //cv::waitKey(0);}// convert tiny_cnn::image to cv::Mat and resizecv::Mat image2mat(image<>& img) { cv::Mat ori(img.height(), img.width(), CV_8U, &img.at(0, 0)); cv::Mat resized; cv::resize(ori, resized, cv::Size(), 3, 3, cv::INTER_AREA); return resized;}void convert_image(const std::string& imagefilename, double minv, double maxv, int w, int h, vec_t& data) { auto img = cv::imread(imagefilename, cv::IMREAD_GRAYSCALE); if (img.data == nullptr) return; // cannot open, or it's not an image cv::Mat_<uint8_t> resized; cv::resize(img, resized, cv::Size(w, h)); // mnist dataset is "white on black", so negate required std::transform(resized.begin(), resized.end(), std::back_inserter(data),  [=](uint8_t c) { return (255 - c) * (maxv - minv) / 255.0 + minv; });}